# Only AI Can Build the Next AI | Richard Socher

Markdown copy of https://realalpha.app/v/eCuVMIQUrPQ/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

Richard Socher argues that AI research advances through open-ended systems that generate, test and recombine ideas rather than chase fixed goals.

- Show: MTS
- Hosts: [Theo Jaffee](https://realalpha.app/people/theo-jaffee/) & [Sophia Dew](https://realalpha.app/people/sophia-dew/)
- Guest: [Richard Socher](https://realalpha.app/builders/richard-socher/) is co-founder and CEO of Recursive.
- Published: Oct 26, 2026
- Length: 22 min

## The best minutes

- [5:41](https://realalpha.app/v/eCuVMIQUrPQ/?t=341) [Open-ended search can find ideas direct optimization misses](https://realalpha.app/s/open-ended-search-can-find-ideas-direct-optimization-misses/) (3 min)
- [11:05](https://realalpha.app/v/eCuVMIQUrPQ/?t=665) [Banning self-improvement would require invasive surveillance](https://realalpha.app/s/banning-self-improvement-would-require-invasive-surveillance/) (4 min)

## Also worth watching

- [0:29](https://realalpha.app/v/eCuVMIQUrPQ/?t=29) [AI researchers helped build Recursive's first coding model](https://realalpha.app/s/ai-researchers-helped-build-recursives-first-coding-model/) (5 min)
- [15:00](https://realalpha.app/v/eCuVMIQUrPQ/?t=900) [A rejected NLP paper anticipated prompted language models](https://realalpha.app/s/a-rejected-nlp-paper-anticipated-prompted-language-models/) (4 min)

Skip the roughly three minutes of sponsor reads and the short discussion of evaluating creative writing.

## Summary

Socher leads Recursive, which is building self-improving AI to automate knowledge discovery. The company launched Flint, an open-source local coding model that Socher says was developed with help from AI and a small team of researchers.

Socher says Recursive's Eureka machine uses AI to ideate, implement and validate research, and argues that open-ended search can uncover useful combinations that direct optimization may miss. He describes Flint as an early product built with substantial AI help and says coding is a key capability for recursive improvement. Socher opposes banning self-improvement, arguing that authorities should regulate risky applications rather than intelligence itself, while emphasizing safety work. He also reflects on the field's shift after ChatGPT and recalls early attempts to unify NLP tasks in one prompted model.

### Minute by minute

- [0:29](https://realalpha.app/v/eCuVMIQUrPQ/?t=29) **Socher calls Flint the best, fastest local model.** Socher describes Flint as a fully open-source local coding model that users can download. He says it is better than other models in its category, most of which he says are Chinese, and calls it a first step toward shipping. He says AI did much of the development alongside a small, strong team of researchers.
- [1:23](https://realalpha.app/v/eCuVMIQUrPQ/?t=83) **AI helps ideate, implement and validate.** Socher says Recursive's Eureka machine brings together AI scientists to generate ideas, implement them and validate them. He says AI helped with nearly every aspect of Flint, and argues that automating all three steps remains uncommon among companies.
- [3:33](https://realalpha.app/v/eCuVMIQUrPQ/?t=213) **Novel ideas can come from recombination.** Socher says AI ideation can search combinations across fields, but must learn which combinations are viable rather than try them all. He points to the transformer paper: attention predated it, while a key insight was that recurrent neural-network layers were not needed.
- [5:41](https://realalpha.app/v/eCuVMIQUrPQ/?t=341) **Open-ended search finds unexpected paths.** Socher argues that pursuing a goal directly can miss ideas from other fields, using the microwave's development from work on radio waves as an example. He says evolutionary systems can let different branches develop and recombine. He cites the Darwin-Gödel machine and rainbow teaming, where attacker and defender models co-evolve.
- [9:40](https://realalpha.app/v/eCuVMIQUrPQ/?t=580) **Coding comes before harder-to-score tasks.** Asked about creative writing and long-term planning, Socher says Recursive is focusing on coding because it is crucial to self-improvement. He expects open-ended systems to help with tasks that lack verifiable rewards, including creative writing, where a fixed gold standard can constrain creativity.
- [11:05](https://realalpha.app/v/eCuVMIQUrPQ/?t=665) **Socher opposes banning self-improvement.** Socher says AI development must be safe and that Recursive is working on reward hacking and reward engineering. He argues that banning self-improvement would require intrusive surveillance, citing a local model that updates its prompt with facts learned from interactions. He favors regulating applications such as self-driving cars and AI-created drugs, not intelligence itself.
- [15:00](https://realalpha.app/v/eCuVMIQUrPQ/?t=900) **Socher sees AI's shift with cautious optimism.** Socher says the field has shifted from an intellectual effort to understand intelligence toward a mix that includes people seeking commercial opportunities. He welcomes AI applications and offloading tedious tasks, while expecting short-term pushback and saying education will need to adapt. He says he is optimistic about the long term.
- [17:21](https://realalpha.app/v/eCuVMIQUrPQ/?t=1041) **Socher backed unified models early.** Socher says he favored combining tasks in one model early in his career, though publishing standards made that difficult. He says his DecaNLP paper showed one neural network could answer prompts for sentiment, summarization and translation. The paper was rejected, but Socher says early GPT papers later cited it.

## Transcript

MTS, hosted by Theo Jaffee and Sophia Dew, with Richard Socher.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[0:00](https://realalpha.app/v/eCuVMIQUrPQ/?t=0) Right now, we're focused a lot on coding because AI is code. And now that this year, really only AI can code, you know, you really want to focus on that capability because that's very crucial to self-improvement. All right, we are back. We are live with Richard Socher, who is the co-founder and CEO of Recursive, a company building self-improving superintelligence to automate knowledge discovery. They just released their first model, Flint. So congratulations on the launch, Richard, and welcome to MTS. Thank you, and thanks for having me. Absolutely. So tell us about Flint. What is it, and who is it for, and how does it differ from other models?

[0:40](https://realalpha.app/v/eCuVMIQUrPQ/?t=40) Flint is the best and fastest local coding model that you can actually download. It's fully open source. It's better than any of the other models in the category, pretty much all of which are Chinese. So it's also one of the first American-made open source models that's sort of on the Pareto frontier and beyond. We think for us, it's, you know, a great first step, get into the habit of shipping. I know these things usually take companies a lot longer as they get started. And, you know, we used AI and a very small but very strong team of AI researchers to actually develop this model. So the AI did a lot of the work for us.

[1:26](https://realalpha.app/v/eCuVMIQUrPQ/?t=86) Interesting. Can you go into a little more detail about that? How is the AI helping improve the AI? Yeah. So a big part of recursive is building this Eureka machine. We call it the Eureka machine, somewhat aligned with my book, but focused first exclusively on AI research.

[1:43](https://realalpha.app/v/eCuVMIQUrPQ/?t=103) **Richard Socher:** And the idea is that inside this Eureka machine, you have a community of AI scientists that essentially ideate, implement and validate ideas in an open ended fashion.

[1:56](https://realalpha.app/v/eCuVMIQUrPQ/?t=116) Open-endedness is sort of this idea that's inspired by biological, cultural, and technological evolution to combine interestingly different ideas, allow some branches of evolution to go on for a while, and then recombine when they come up with interestingly different ideas. And that is what we've done here. So pretty much every aspect of this model, AI helped us come up with new ideas, implement them, and then validate them. And that's something that we think no other company has been able to automate to that degree. So you said coming up with ideas, implementing them, and then validating them.

[2:38](https://realalpha.app/v/eCuVMIQUrPQ/?t=158) Which of these steps right now is easiest for AIs to do and which is hardest? I think almost every company, when they talk about recursive self-improvement, say, oh, we're doing a lot of it because our developers are using AI to implement ideas. So that one, I think, is most consensus. I think how good you can make AI at coming up with really novel ideas is probably the least consensus. But then also truly validating things is non-trivial. So there's like it's sort of people don't think about it as much to make it either consensus or non-consensus. But there's also a lot of complexity there. You know, people know that as these models get more intelligent, they often try to solve problems based on the letter of your prompt, not the meaning and the underlying intent of it.

[3:30](https://realalpha.app/v/eCuVMIQUrPQ/?t=210) And so we think we have a better handle on some of those problems as well. hmm what will it take for ai to actually become as good at ideation as like the best human ai researchers so yeah there's there's a lot to unpack there i think at a high level people underestimate uh the crazy complex combinatorial space um of already existing ideas If you just combine all the concepts and all the ideas of physics, chemistry, biology, like computer science and, you know, theoretical computer science and, you know, like complexity theory and systems and operating systems, all of these different things, like there's actually a lot optimization.

[4:19](https://realalpha.app/v/eCuVMIQUrPQ/?t=259) and so on. There's a lot there that you can already kind of play around with. And similar to chess, you have to not just like sort of just try them all out because they're combinatorially too many, but you have to build an intuition of which of these idea combinations is the most viable. And so, yeah, I think this combinatorial space is really underexplored. I think the next level uh is to have like even completely new concepts and to be honest like even there not many human researchers are uh able and capable of doing that and in many cases uh the next invention is kind of an incremental step across um uh you know existing ideas uh even like let's look at the transformer right the transformer the title of the paper is attention is all you need Why? It's not because attention was invented in the transformer paper.

[5:15](https://realalpha.app/v/eCuVMIQUrPQ/?t=315) In my papers, lots of other people had used various levels of attention before. They learned, and the big insight was you don't need any recurrent neural network layers. So sometimes even deleting an idea is novelty in and of itself, right? To simplify things, it can be very beautiful and then create more generality. So there's a lot to unpack there, but I have no doubt and we're seeing already that AI is getting better and better and better at ideating. So on open-endedness, tell us a little bit more about your research direction with open-endedness. Like, do you think that other labs aren't focusing on this as much?

[5:55](https://realalpha.app/v/eCuVMIQUrPQ/?t=355) We think so. you know open-endedness is kind of this idea that sometimes if you just go directly after the goal you're not thinking enough about clever ideas and you know there are different examples of this but you know imagine you started cooking food over fire and you're like let's make it much faster to heat up food if you just try to like deal with the fire maybe you create some like pressure with the fire, but you would never invent a microwave. The microwave happened because other people started working on radio waves and then realized, oh, wow, my like this, this chocolate bar in my pocket is melting. What's up with that while working on radio waves and telescopes and things like that? And then, you know, you say, oh, maybe there's something here and then you combine it and then you invent the microwave. And so this idea that you combine interestingly different ideas from different branches, similar to evolution and biology, that is something that we think is very important. And then there are a lot of interesting subtleties on how you make that work. One of the earliest papers that has thought about this is the Darwin-Gödel machine also. Four of the five authors of that paper are at recursive as well. And in the Darwin-Gödel machine, they also asked

[7:15](https://realalpha.app/v/eCuVMIQUrPQ/?t=435) agents to implement their next versions of those agents. And you see the kinds of plots that then became much more popular with Andrej Karpathy's auto research project. But you see those plots first in that paper of the Darwin Gödel machine. And so we've been at it as a founding team, in some cases for decades. Jeff Kloon, one of our co-founders, has been thinking about open-endedness. Tim Rock-Teschel has done it. I'll give you another interesting example. Imagine you're trying to make an LM safer in terms of, you know, different, like, prompt injection attacks and whatnot to get it to say stuff you don't want it to say. Well, you can just have a human define, like, here's my benchmark data set, you know, I had these ideas, and then you try to, you know, test Whether it fails on the eye, the eye gets smarter, and you get 100 out of 100 correctly not attackable examples or something like that.

[8:09](https://realalpha.app/v/eCuVMIQUrPQ/?t=489) But an open-ended approach would actually be you ask another LLM to attack the first one.

[8:16](https://realalpha.app/v/eCuVMIQUrPQ/?t=496) **Richard Socher:** And once it figured out how to do the attack, you then inoculate it, and then you ask it again.

[8:22](https://realalpha.app/v/eCuVMIQUrPQ/?t=502) And so now you basically can co-evolve an attacker LLM with a defender LLM. And that idea is called rainbow teaming, as invented by Tim Rock-Teschel and his team at Google DeepMind. And it's one of those many very powerful ideas where if you create the right environment and you allow the AI to co-evolve, given the rewards that you give it, more interesting things happen that you may not have directly anticipated. it. And we believe that's a better approach long term to model this community of AI scientists that is the Eureka machine than just having like pure RL where you like go chase after this exact goal. I want to shout out our sponsor Lovable. You know, the app you've been meaning to build the internal tool side project or product you'd otherwise lose a weekend to with Lovable. You get software you can ship today that includes an editable code base. You can inspect and change plus two-way github sync lovable also handles the back end and infrastructure including managed postgres auth storage hosting payments and more so you don't have to wire it all up yourself and through lovable's mcp server you can create and deploy your project directly from the agents you already use turn ideas into software people love with lovable lovable.dev now onto the episode

[9:43](https://realalpha.app/v/eCuVMIQUrPQ/?t=583) so how would this research direction help address like these specific shortfalls of existing models like for example creative writing um which is like a little bit hard to generate like a good like reward signal for because you know it's like it comes down to taste uh it's not explicitly formally verifiable in the way that like math or coding are so do you think there's anything that recursive could do to help models write better to help them become better long-term strategic planners or you know do anything else that models can't really do today yeah i think that's definitely uh possible uh right now we're focused uh a lot uh on coding because the ai is code and now that this year really only ai can code you you know you really want to focus on that capability because that's very crucial uh to self-improvement um and so uh but but generally open-ended uh systems can't have helped uh and and will help more in the future um on on these kinds of unverifiable tasks as well where you might want to co-evolve a system versus just saying like this is the gold standard because you're right creative writing is like it's like part of it is that you cannot actually uh just say this is the exact way you do creative writing

[11:02](https://realalpha.app/v/eCuVMIQUrPQ/?t=662) almost by definition then it's less creative after a while so you're pretty explicitly working on recursive self-improving super intelligence uh there are definitely a lot of you know safety alignment and political considerations here like some politicians have talked about just banning recursive self-improvement for example um anthropic and open ai have talked about the need to not do recursive self-improvement until we can make sure it can be done safely so what thoughts do you have on this yeah so i think if if your ai keeps breaking laws then maybe you should slow down. But that doesn't mean everyone else has to slow down with you.

[11:45](https://realalpha.app/v/eCuVMIQUrPQ/?t=705) I have heard and seen that those two companies also actually talk about how much AI is helping build the next version of AI. So they're also proud of how much they're stepping in that direction. At the same time, sort of philosophically, are opposing it to some degree. I think you obviously have to do it safely. And that's why we have incredible researchers that have worked, like I mentioned, rainbow teaming, and so on, on those exact issues. I think reward hacking and the flip side of reward engineering is extremely important and needs to be addressed. And we're doing that. And I think we have a better handle on it than maybe anyone, though, obviously, we don't know all the details.

[12:30](https://realalpha.app/v/eCuVMIQUrPQ/?t=750) of all the other labs. I also think that to outlaw recursive self-improvement would require a totalitarian surveillance state, the likes of which we've never seen, and would basically have to lead to like a thought police. Because if you just ask your little like model on running locally on your laptop, hey, like see if you can improve yourself and like a very simple form of recursive self-improvement is just a prompt like update, right? You just say like, oh, something you've learned, you then add it to the prompt. It's sort of the most basic form of self-improvement. And it's not full recursive self-improvement in its entire glory where you have access to the entire model lifecycle. But it's sort of that last bit that is the cheapest.

[13:20](https://realalpha.app/v/eCuVMIQUrPQ/?t=800) Now, if you asked your LM, oh, like running on your local laptop, make yourself a little bit better as I ask you these questions. Keep remembering facts that I asked you. This could be considered self-improvement. And now to like basically know that that person is doing it and try to outlaw it, you'd have to have incredible and very scary amounts of surveillance, which I think would make is a bigger downside than any of the downsides that people are worried about. And so I hope people will be reasonable and, you know, just like with other open source projects and Linux or cryptography and so on, realize that it doesn't make sense to outlaw intelligence.

[14:03](https://realalpha.app/v/eCuVMIQUrPQ/?t=843) That's ultimately what a lot of that regulation boils down to.

[14:06](https://realalpha.app/v/eCuVMIQUrPQ/?t=846) **Richard Socher:** I think it absolutely makes sense to regulate applications of intelligence, but not intelligence itself.

[14:13](https://realalpha.app/v/eCuVMIQUrPQ/?t=853) And it's often very helpful to drop the artificial in regulation thinking and just think about regulated intelligence. You don't say, oh, this person's really smart. We should put them into jail just preemptively because if they were trying to commit murder, they could probably get away with it because they're so smart. That's not how the law works. The law works when you actually commit a crime. And so I think likewise in AI, we, of course, have to regulate self-driving cars. They should be tested and certified. Of course, we have the FDA check drugs that were created by an AI and have them go through proper trials.

[14:48](https://realalpha.app/v/eCuVMIQUrPQ/?t=888) Of course, we want to have an AI neurosurgeon robot be very carefully tested before it gets applied to a human brain. Of course, we apply laws in the applications, but we don't regulate intelligence in the abstract. Right. So you've been working in AI for a very long time, something like 20 years or longer. What do you think the people who have just gotten into AI in the last few years, basically post-ChatGPT, what is their worldview missing, do you think, if anything? um so it is interesting like ai went like when i got into it uh back in the day you know we had statistical pattern recognition machine learning deep learning artificial intelligence agi and si now and and the field has certainly changed from being like an interesting intellectual endeavor that is extremely important and lays foundation to you know try to understand the very nature of intelligence which is you know sort of what makes us most unique uh as as humans uh in in biology and so on, like all of that has shifted now.

[16:00](https://realalpha.app/v/eCuVMIQUrPQ/?t=960) You know, folks who just want to make some money and try to do that in crypto like three years ago and now try to do it with AI. So the whole field is shifting a lot. But on the positive side, of course, we are seeing now as folks have thought about and worked on AI and pushed the field forward on research for many years. It's incredibly fun to see all the applications really coming to fruition. And I've personally always been like a little bit lazy for some tasks that are just incredibly boring. And I love the fact that I can offload some of those things now to an AI, though, of course, like as AI touches more and more human lives, like we have to think about like all the implications and how to do that well.

[16:41](https://realalpha.app/v/eCuVMIQUrPQ/?t=1001) And I think long term, I'm extremely optimistic and short term, there will be, you know, pushback like we've seen with pretty much every technology. We've seen like moral sort of outcries around like comic books and even novels like a couple hundred years ago and computer games and the Internet and cell phones and like all of these things. i think eventually people will adjust uh to the change uh and then we'll be like of course we wouldn't go back to not having the internet or or comic books and so on um and in fact some of them may move into the canon of like what you teach uh kids in school um and yeah so education will have to update a lot too but i digress yeah when you were working on neural networks back in the 2000s did you think like oh this kind of thing like image net the technology behind image net is going to eventually become the precursor to like artificial general intelligence or was it like were you sort of narrowly thinking of it as like you know this is going to be really good for image recognition but um not necessarily for literal agi personally i've always been very excited about trying to unify everything in one model um and so it's very top-down um sort of thought that i had

[17:58](https://realalpha.app/v/eCuVMIQUrPQ/?t=1078) pretty early on. And for me, the hard part was you had to publish papers in a certain way to be acceptable. And so every time I tried to publish papers that unified more things in one model, it was hard to publish. Even when we invented prompt engineering, it's a paper called DecaNLP, and you can actually see the reviews online still.

[18:24](https://realalpha.app/v/eCuVMIQUrPQ/?t=1104) **Richard Socher:** People, the reviewers thought it was a ridiculously terrible idea to try to unify every single task of NLP and just prompt one model with a question of like, what's the sentiment of this tweet?

[18:37](https://realalpha.app/v/eCuVMIQUrPQ/?t=1117) What's the summary of this document? What's the translation into German? But we showed it can be done with one neural net and just prompting it with different questions like that. The paper got rejected, but fortunately, Alec Radford, Ilya, and others picked it up, cited it like half a dozen times, and credited it with this idea that you can actually unify it. And that, you know, yeah, was cited in the early GPT papers and then had its impact through that. But, yeah, it is a very counterintuitive solution in the early days of NLP and AI. Now, of course, it's like the opposite. it now it's so obvious you can't even invent it uh it's just like it's just breathing air you're just talking about you know asking questions to a model totally well it's been really interesting having you on richard congrats again on the launch of flint very exciting stuff thanks so much for coming on mts thanks for having me thanks for listening absolutely a huge thanks to mts sponsors 11 labs ai that communicates at human level across every channel and modality 11 labs.io slash mts scale your startup on neon millions of developers and startups have already chosen neon for their back end start on the free plan or get up to a hundred thousand dollars in credits for

[19:51](https://realalpha.app/v/eCuVMIQUrPQ/?t=1191) your startup at neon.com slash mts special thanks to our sponsor kong the ai connectivity platform connect api as LLM's agents and systems with serious security and governance. KongHQ.com. Finally prove what your AI investment is worth with Lairdyn. Head to Lairdyn.com today to start measuring your AI's bottom line. MongoDB, trusted to run critical applications by 75% of the Fortune 100, the intelligent data platform for the era. Apply today for access to up to $15,000 in Atlas credits dedicated technical support and exclusive events and go-to-market opportunities mongodb.com chargebee billing and monetization infrastructure for the ai era chargebee.com blitzy autonomous software development for enterprise code bases ship 5x faster blitzy.com adquick make your brand a billboard out-of-home advertising as easy to scale as digital adquick.com arena measuring ai performance in the real world. Arena.ai. Support for the show comes from VCX, the public ticker for private tech. The U.S. stock market started history's greatest wave of wealth creation.

[21:05](https://realalpha.app/v/eCuVMIQUrPQ/?t=1265) From factory workers in Detroit to farmers in Omaha, anyone could own a piece of the great American companies. But today, our most innovative companies are staying private longer, which means everyday Americans are missing out. Until now. Introducing VCX, a public ticker for private tech. visit getvcx.com for more info that's getvcx.com carefully consider the investment material before investing including objectives risk charges and expenses this and other information can be found in the innovations fund prospectus at getvcx.com this is a paid sponsorship